JPS58176698A - Pattern matching system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention The present invention relates to a pattern matching method, and more particularly, to a pattern matching method suitable for a variety of applications requiring high-speed execution, such as a vocal pattern recognition device.

Conventional pattern recognition devices perform pattern matching between human cover turns and standard patterns, but voice pattern recognition devices and the like need to perform this at high speed. Such fast pattern matching is generally not fast enough to perform program calculations using a general-purpose arithmetic unit such as a microcomputer. Therefore, conventional f/'' pattern recognition devices are configured to perform pattern matching using a suitable, dedicated high-speed arithmetic unit.However, in order to improve recognition accuracy, pattern matching calculations are When trying to improve accuracy,
In such a configuration, the hardware cost is likely to increase, leading to a significant increase in the price of the device.

OBJECTS OF THE INVENTION It is an object of the present invention to provide a pattern matching method for a pattern recognition device that achieves the same recognition accuracy while suppressing an increase in domain cost as much as possible.

Although the program operation using the general purpose arithmetic means (2) of the microcomputer is slower than the dedicated arithmetic means by Jl, it is easy to reverse this; it is possible to obtain a high arithmetic ffK.

The present invention uses this dotted layer, and performs pattern matching between the input cover pattern and the standard pattern by performing two types of pattern matching: primary pattern matching using a dedicated calculation means, and 2(′) pattern matching using a general purpose calculation means. In the first-order pattern matching, the human cover turns are identified by performing the second-order pattern matching with increased calculation accuracy only for those that cannot be identified with a sufficient number of 6[.

Secondary pattern matching should not be performed on all standard patterns, but only on a limited number of standard patterns including the first and second complementary patterns extracted in primary pattern matching. This aims to shorten the execution time.

Embodiments of the Invention Two embodiments of the present invention will be described with reference to the accompanying drawings.

In this embodiment, the matching distances to the human cover pattern are computed at high speed for all standard patterns by a dedicated computing means. Among the standard patterns whose matching distance is equal to or greater than the specified value, the standard pattern with the minimum matching distance is used as the first complementary pattern, and the next smaller standard pattern is used as the second complementary pattern.
Selected as a candidate pattern. Then, if the difference in the matching distance between the first and second candidate patterns is equal to the specified value and is F+, the human cover turn is compared to the first candidate pattern, and the recognition ends. However, if the matching end of the first and second box patterns is less than the specified value, a limited number of candidate patterns including the first and second candidate patterns can be calculated using a general-purpose calculation means. standard pattern (only two calculations with a higher matching distance with the human cover turn than the dedicated calculation means described above)
It is recalculated at the same time.

Then, as a result of this recalculation, the above judgment is performed for the two groups.

Each part of laF and Honji Mfflj will be explained in detail.

This embodiment is a device for recognizing voice patterns, and voice is converted into an electrical signal by a microphone 1 and input. This audio signal is input in parallel to a group of bandpass filters 31 (i=1.2°...+n) via a buffer amplifier, and the audio frequency components of, for example, 200Hz to 6000Hz are frequency-analyzed into n channels. be done. The number of channels n is typically 16 to 16 channels.

The output signal of the bandpass filter 3i has a time constant of 50
A group of low-pass filters 4i (t=t,
2. . . t") in accordance with the channel, and the envelope waveform of the back door waveform is converted. Low-pass filter 4
Each output signal of 1 is generated approximately every 4 milliseconds (1 frame) I:,
An analog/digital (A/D) converter 6 is connected via an analog switch 5 in a time-sharing manner under the control of a control unit 90.
sent to. - A/D conversion a6 of 8 to 12 bits is used, and its output signal is temporarily stored in the buffer memory 7 in the form of one frame of sperature information about every additional millisecond for each channel.

The spectrum Fft information temporarily stored in the buffer memory 7 is normalized by performing the calculation shown in equation (1), for example.

Here, Pi is the power output of channel i, n is the total number of channels, and C is a constant. That is, according to this equation (1), ζ-2 normalization is performed so that the sum of the outputs of each channel after normalization becomes equal to the constant C.

This normalization operation is executed by a general purpose arithmetic unit 80 microprogram configured by a microcomputer.

In addition, the i-voice section can also be detected under the monument ``General-Purpose Arithmetic Unit 80 Microprogram (Second, Control Unit 90 City 1'').
Two runs. Various algorithms are known, but for example, frames with a total power of a predetermined value of 1 or more are considered to be non-audio signal frames, and frames with less than a threshold are considered to be noise signal frames. Then, when there are five or more consecutive frames of the audio signal, extraction is started as a voice section, and thereafter, when frames of the noise signal are consecutive on IOα, the voice section is detected by an algorithm that determines the end of the voice section.

Note that information on a series of frames cut out as speech sections (speech pattern information) is stored in the RAM (random access memory) 10 when registering standard patterns, but when recognizing human speech (the second is DP Matsutengu Units). 11
sent to. Note that the registration of the porosity pattern is not the gist of the present invention and may be similar to the conventional apparatus, so the details will be omitted.

1) P matching unit) 11 is a general-purpose calculation unit 8
The system performs pattern matching (referred to as first-order pattern matching) based on so-called DP (dynamic programming) between the input audio pattern sent from the input audio pattern and the audio pattern (standard pattern) stored in RAM10. Rabbit running with luck. It is a dedicated special calculation unit. Note that, as is well known, the DP matching method can absorb variations in the generation speed of standard patterns and human speech patterns, and is recognized as being effective as a pattern matching method in speech recognition. Further, the DP matching unit 11 may have a known configuration.

In the DP matching unit 11, matching distances between each standard pattern and the input audio pattern are sequentially calculated and sent to the determination unit 12. The determination unit H2 selects the standard pattern with the smallest matching distance (highest similarity to the human voice pattern) as the first candidate pattern among the standard patterns whose matching distance is less than or equal to a specified threshold value, and the second candidate pattern with the smallest matching distance. Select the standard pattern as the second candidate pattern. If the difference in matching distance between the first and second candidate patterns is equal to or greater than a specified value, the first candidate pattern is selected as the recognition result. If this condition is not met, the microprogram of the general-purpose arithmetic unit 8 increases the arithmetic precision and performs DP again.
Matching (referred to as secondary pattern matching) is performed. Then, when the result of this secondary pattern matching is 1'', the determination unit 12 makes a similar determination.If the above conditions are still not satisfied, the input speech pattern is rejected as unrecognizable. Ru.

The calculation speed of the general-purpose calculation unit 8 is significantly slower than that of the dedicated ha unit l)P matching unit) 11. Therefore, if we execute the second-order pattern matching for all standard patterns, the execution time will be
It's too much. Therefore, in the present invention, secondary pattern matching is performed only on a limited number of standard patterns including the first and second candidate patterns for primary pattern matching, thereby reducing the execution time. . The standard pattern used for acid resistance at this time is DP matching unit) 1
1 (can be easily selected by the method of second-order matching, etc.).

Here, the performance of the matching distance in the DP matching method J!
Explain accuracy.

Let the value of the normal 1 Hibwerworth vector of human force fF expressed by the above equation (1) at time t be NPl(t), and similarly, the value of the normalized spectrum of #4 quasi-pattern at time τ be N5Pi(τ). Then, DP matching is performed to calculate the matching distance between the two (the two-dimensional interlattice distance is expressed, for example, by Equation (2)).

,Σ(NPl(t)−N8Pi(τ))2...
... Equation (2) The smaller the inter-lattice distance, the more similar the normalized power spectrum at time t (2) of the input voice and the normalized power spectrum at time τ of the standard pattern are. Since the distance is the basis for calculating the matching distance, its functional form has been devised in various ways. For example, a linear equation such as equation (3) has been considered instead of equation (2).

i mi, l NPi(t)-N,5Pi(τ)1 ・
・・・・・・Equation (3) Such a calculation of the interlattice distance A quadratic equation like equation (2) emphasizes the difference in the normalized power spectrum, so a quadratic equation like equation (3)
In the case of the following equation: When comparing the two, when DP matching is performed to find the vertical@- distance, the characteristics of the power spectrum corresponding to the bolt mant of the i voice are fully reflected by 1. Therefore, the matching distance between i-voices of the same utterance content (same category) becomes small, whereas the matching distance between i-voices of different categories becomes significantly large, and as a result, the determination is made for each station.

Therefore, in order to obtain the matching distance, it is preferable that the DP matching unit 11 adopts a quadratic equation such as equation (2). The entire process must be performed using an arithmetic unit with a large number of digits; an increase in the hardware cost of the DP matching unit 11 is unavoidable. It is conceivable to truncate the digits of
Another possible method is to set an upper limit value for the value of equation (2) and replace it with the upper limit value when the value exceeds the upper limit value. However, in this case, the spectral emphasis is overwritten, and as a result, the difference in the matching distance between Ondo tiles of different categories is reduced, which is the same as reducing the calculation accuracy.

As described above, if recognition accuracy is to be achieved only by pattern matching using the DP matching unit 11, which is a dedicated high-speed calculation means, a significant increase in hardware cost is unavoidable.

According to the present invention, first-order pattern matching is performed using a dedicated high-speed calculation means, and if sufficient determination is not possible with that, a second-order pattern matching with high calculation accuracy is performed using a general-purpose calculation means. Perform matching.

Therefore, recognition accuracy can be ensured without much restriction on the calculation accuracy of the primary pattern matching, and the hardware cost of the L)P matching unit 11, which is a dedicated calculation means, can be reduced. In addition, the general-purpose arithmetic unit 8, which is a general-purpose arithmetic means, executes a program (secondary calculation), so the calculation speed is limited. However, only when the first-order pattern matching cannot determine the
In addition, the second-order pattern matching is performed not on all standard patterns, but on a limited number of standard patterns including the standard patterns extracted as candidate patterns in the first-order multiple-turn matching. Executes only standard turn. Therefore, the increase in the execution time of pattern matching can be kept to a very small level, and a sufficient degree of recognition can be achieved. Can be done.

The invention has been described in detail with respect to the example of a speech pattern recognition device.

However, the present invention is not limited to the acupuncture device that recognizes the voice pattern. It goes without saying that the present invention can be similarly applied to other pattern recognition devices such as a computer.

According to the present invention, high-definition sieving speed pattern matching is possible while suppressing the increase in hardware cost for pattern matching as much as possible. As a result, it is possible to realize the pattern of a first-year high school student, who has a high recognition rate and recognition speed, in just one song (or two) using a ninja device.

[Brief explanation of drawings]

The figure is a block diagram showing one embodiment of the present invention. 1...Microphone, 2...Buffer booster, 3i
(31-3n)...Banthus filter, 4i (4
t~4n)...Low pass filter, 5...Analog switch, 6...Analog/digital converter, 7.
...Buffer memory, 8...General purpose arithmetic unit, 10
... Random access memory, 11...l) P matching unit l-, i2... Judgment unit.

Claims (1)

[Claims] (1) Pattern recognition! It! [1-, the pattern matching between the input cover pattern and the standard pattern is divided into primary pattern matching using dedicated calculation means and secondary pattern matching using program calculation using general-purpose calculation means. , the first-order pattern matching is executed for all standard patterns, and the difference in the distance between the two standard patterns selected as candidate patterns corresponding to the incoming cover turn in the first-order pattern matching with the human cover turn. 2 as the candidate pattern only when is less than the specified value.
A pattern matching method characterized in that the 4M2 pattern matching is executed with higher calculation accuracy than the first pattern matching only for a limited number of standard patterns including two absolute patterns.